Development and validation of radiomics models for the diagnoses of breast lesions with calcification on ultrasound.

Background: High-frequency ultrasound (US) technology has enhanced the visualization of breast calcification; however, the understanding of these calcifications among sonographers remains limited. This study aimed to investigate the diagnostic potential of US images for breast lesions with calcification via a radiomics approach. Methods: This retrospective study analyzed breast lesions with calcification (verified on mammography) at a single center. Additionally, 56 lesions from the public Breast Ultrasound Image Dataset (BUSI) served as the external test set. Regions of interest (ROIs) for lesions and calcifications were delineated on the US image respectively and used for feature extraction. Pathological findings and follow-up assessments were used as the diagnostic criteria for benign and malignant lesions. Radiomics models were developed using lesion, calcification, and fusion features based on six machine learning algorithms. In addition to the receiver operating characteristic curve and other metrics, a reader study was performed to evaluate the model performance. Results: A total of 508 US images of breast lesions from 490 women (mean age, 52.1±11.0 years) were included. Among the 919 extracted two-dimensional radiomics features, 44, 14, and 23 features were selected to construct the calcification-based, lesion-based, and fusion-based models, respectively. The fusion strategy outperformed the other two strategies across most algorithms and exhibited enhanced diagnostic performance on both test sets. In terms of the area under the receiver operating characteristic curve (AUC), the fusion models significantly outperformed the calcification-based models for the Decision Tree (DT) and Random Forest (RF) algorithms on the internal test set (both P<0.05), and significantly outperformed the lesion-based models for the Bootstrap Aggregation (BAG) and Logistic Regression (LR) algorithms on the external test set (both P<0.05). Notably, the AUC of the representative BAG model improved significantly after incorporating calcification features, increasing from 0.679 to 0.770 on the external test set (P<0.05), which was comparable to the performance of most experienced sonographers. Conclusions: Fusing features of lesion and calcification on US images enhance the predictive efficacy of radiomics models, indicating the additional value of calcification in diagnosing breast lesions.